Variable Importance from mlr3

Data prep

data("Servo")
servo <- Servo %>%
  mutate_at(c("Pgain", "Vgain"), as.character) %>%
  mutate_at(c("Pgain", "Vgain"), as.numeric)
head(servo)

##   Motor Screw Pgain Vgain Class
## 1     E     E     5     4     4
## 2     B     D     6     5    11
## 3     D     D     4     3     6
## 4     B     A     3     2    48
## 5     D     B     6     5     6
## 6     E     C     4     3    20

train_size <- 2/3
set.seed(1333)
train_index <- sample(
  x = seq(1, nrow(servo), by = 1),
  size = ceiling(train_size * nrow(servo)), replace = FALSE
)
train_1 <- servo[ train_index, ]
test_1 <- servo[ -train_index, ]

task <- TaskRegr$new(id = "servo", backend = train_1, target = "Class")
lrn1 <- lrn("regr.ranger", importance = "impurity")
lrn1$train(task = task)
filter <- mlr3filters::flt("importance", learner = lrn1)
filter$calculate(task)
var <- as.data.table(filter)
ggplot(data = var, aes(x = feature, y = score)) + geom_bar(stat = "identity") +
  ggtitle(label = "Variable Importance with mlr3") +
  labs(x = "Feature", y = "Variable Importance")

spiral <- mlbench::mlbench.spirals(1000, cycles = 2, sd = 0.5)
p <- ggplot(data = as.data.frame(spiral$x), aes(
  x = V1, y = V2,
  colour = spiral$classes
)) +
  geom_point()
p

Decision regions CART

spiral_data <- data.frame(spiral$x, y = factor(spiral$classes))
colnames(spiral_data) <- c("x1", "x2", "y")

features <- c("x1", "x2")
spiral_task <- TaskClassif$new(
  id = "spirals", backend = spiral_data,
  target = "y"
)
plot_learner_prediction(
  lrn("classif.rpart", predict_type = "prob"),
  spiral_task
)

## INFO  [17:14:32.188] [mlr3] Applying learner 'classif.rpart' on task 'spirals' (iter 1/1)

Decision regions Random Forest

plot_learner_prediction(
  lrn("classif.ranger", predict_type = "prob"),
  spiral_task
)

## INFO  [17:14:33.402] [mlr3] Applying learner 'classif.ranger' on task 'spirals' (iter 1/1)

set.seed(1337)
spiral_rf <- randomForest(
  x = spiral$x, y = spiral$classes,
  ntree = 1000,
  proximity = TRUE, oob.prox = TRUE,
)
spiral_proximity <- spiral_rf$proximity
spiral_proximity[1:5, 1:5]

##           [,1]      [,2] [,3] [,4] [,5]
## [1,] 1.0000000 0.0078125    0    0    0
## [2,] 0.0078125 1.0000000    0    0    0
## [3,] 0.0000000 0.0000000    1    0    0
## [4,] 0.0000000 0.0000000    0    1    0
## [5,] 0.0000000 0.0000000    0    0    1

proximity_to_dist <- function(proximity) {
  1 - proximity
}
spiral_dist <- proximity_to_dist(spiral_proximity)
spiral_dist[1:5, 1:5]

##           [,1]      [,2] [,3] [,4] [,5]
## [1,] 0.0000000 0.9921875    1    1    1
## [2,] 0.9921875 0.0000000    1    1    1
## [3,] 1.0000000 1.0000000    0    1    1
## [4,] 1.0000000 1.0000000    1    0    1
## [5,] 1.0000000 1.0000000    1    1    0

spiral_mds <- as.data.frame(cmdscale(spiral_dist))
spiral_mds$class <- spiral$classes
# plot the result, sweet
plot <- ggplot(data = spiral_mds, aes(x = V1, y = V2, colour = class)) +
  geom_point() +
  labs(
    x = "1st dimension", y = "2nd dimension",
    title = "Multidimensional scaling based on inverse proximity (spirals data)"
  )+
  scale_colour_viridis_d()
plot